BayeSeg: Bayesian modeling for medical image segmentation with interpretable generalizability

Area of Science:

• Medical image analysis
• Deep learning
• Computational imaging

Background:

• Deep learning segmentation methods struggle with cross-domain distribution shifts in medical imaging.
• Existing domain generalization techniques lack interpretability.
• Real-world applicability of current methods is limited by poor performance on unseen data.

Purpose of the Study:

• To propose an interpretable Bayesian framework (BayeSeg) for enhanced medical image segmentation generalizability.
• To address the challenge of interpretability in domain-invariant feature extraction.
• To improve the reliability and applicability of deep learning models in diverse medical imaging scenarios.

Main Methods:

• Developed a Bayesian framework (BayeSeg) using Bayesian modeling of image and label statistics.
• Decomposed images into spatial-correlated (shape) and spatial-variant (appearance) variables with hierarchical Bayesian priors.
• Modeled segmentation as a locally smooth variable related to shape and used a variational Bayesian framework for inference.

Main Results:

• Demonstrated the effectiveness of BayeSeg on prostate and cardiac segmentation tasks through quantitative and qualitative results.
• Validated the interpretability of the framework by explaining inferred posterior distributions.
• Identified factors influencing generalization ability via ablation studies.

Conclusions:

• BayeSeg offers an interpretable approach to enhance medical image segmentation generalizability.
• The proposed framework effectively handles domain shifts by separating shape and appearance modeling.
• The method shows significant potential for improving the robustness and applicability of deep learning in medical imaging.